Appears in Joint Statistical Meetings (JSM), Seattle, 2006
Using Statistics to Search and Annotate Pictures: an Evaluation of Semantic Image Annotation
and Retrieval on Large Databases
Antoni B. Chan1 , Pedro J. Moreno2, Nuno Vasconcelos1
Department of Electrical and Computer Engineering, University of California, San Diego 1
Google, Inc.2
Abstract
of the concept. This has two limitations. First, the complexity of learning each binary classifier depends on the size of the
“absence” class, making the model very difficult to learn when
there is a large vocabulary. Second, the binary formulation is
not amenable to weakly labeled data, in the sense that training
images that contain the concept, but that are mislabeled as not
containing it, are incorrectly assigned to the “absence” class,
and can compromise the classification accuracy.
We present the results of an extensive experimental evaluation of the supervised multi-class labeling (SML) model for
semantic image annotation proposed by [16]. We test the robustness of this model to various parameters, and its scalability in both database and vocabulary size. The results of
this study complement previous evaluations by [12], and [16],
which were limited to smaller databases and vocabularies. We
further compare the performance of SML to that of a model
that explicitly trades off retrieval performance for scalability:
the supervised category-based labeling (SCBL) model of [14].
This establishes a unifying view of the performance of two
classes of labeling and retrieval systems that were previously
only evaluated under different experimental protocols. This
unification simplifies the evaluation of future proposals for semantic labeling and retrieval.
In response to some of these limitations, namely the lack
of scalability to problems involving large vocabularies (hundreds or even thousands of concepts), there has recently been
a shift to the adoption of semantic labeling techniques based
on unsupervised learning [8, 9, 10, 11, 12, 13]. The basic idea
is to introduce a graphical model containing a set of latent
variables that encode hidden states of the world, where each
state defines a joint distribution on the space of semantic keywords and image appearance descriptors (in the form of local
Keywords: semantic image annotation, semantic image re- features computed over image neighborhoods). During training, the image is represented as a collection of regions (either
trieval, experimentation, evaluation.
block-based [10, 12, 14] or segmented regions [8, 9, 11, 15]),
and an unsupervised learning algorithm is run over the en1 Introduction
tire database to estimate the parameters of the joint density of
words and visual features. Given a new image to annotate, viContent-based image retrieval, the problem of searching large
sual feature vectors are extracted, the joint probability model
image repositories according to their content, has been the
is instantiated with those feature vectors, state variables are
subject of substantial research in the recent past [1]. While
marginalized, and a search for the set of labels that maximize
early retrieval architectures were based on the query-bythe joint density of text and appearance is carried out.
example paradigm, which formulates image retrieval as the
search for the best database match to a user-provided query
While the unsupervised formulation leads to significantly
image, it was quickly realized that the design of fully func- more scalable (in database size and number of concepts) traintional retrieval systems would require support for semantic ing procedures and places weaker demands on the quality
queries [2]. These are systems where the database of images of the manual annotations for training, it does not explicitly
are annotated with semantic keywords, enabling the user to treat semantics as image classes and, therefore, provides litspecify the query through a natural language description of tle guarantees that the semantic annotations are optimal in a
the visual concepts of interest. This realization, combined recognition or retrieval sense. That is, instead of annotations
with the cost of manual image labeling, generated significant that achieve the smallest probability of retrieval error, it siminterest in the problem of automatically extracting semantic ply produces the ones that have largest joint likelihood under
descriptors from images.
the assumed model. This has motivated the recent introducThe earliest efforts in the area were directed to the reliable tion of a model, supervised multi-class labeling (SML), that
extraction of specific semantics, e.g. differentiating indoor combines the advantages of both supervised and unsupervised
from outdoor scenes [3], cities from landscapes [4], and de- methods of semantic annotation, through a slight reformulatecting trees [5], horses [6], or buildings [7], among others. tion of the supervised learning model (see [16, 17]). This
These efforts posed the problem of semantics extraction as one reformulation consists of defining a multi-class classification
of supervised learning: a set of training images with and with- problem where each of the semantic concepts of interest deout the concept of interest was collected and a binary one-vs- fines an image class, and has various important practical conall classifier was trained to detect the concept of interest. The sequences. For example, it eliminates the need to compute a
classifier was then applied to all database images which were, “non-class” model for each of the semantic concepts of interin this way, annotated with respect to the presence or absence est, while producing a natural ordering of semantic keywords
1
(i.e., image classes) when annotating a test image. In result,
the new model has learning complexity equivalent to that of
the unsupervised formulation but, like the supervised one, it
places weak requirements on the quality of manual labels. Finally, by explicitly treating semantics as image classes, it provides optimality guarantees in a recognition or retrieval sense.
Independently of algorithmic advances, one of the most significant recent contributions to the advance of semantic labeling and retrieval was the establishment, by [12], of an evaluation protocol that allows easy comparison of any new retrieval algorithm to most of the state-of-the-art techniques.
In particular, [12] proposed a database (which we refer to
as Corel5K), and a set of evaluation metrics (which we discuss in detail in Section 3), and evaluated a number of techniques [11, 12, 15] according to these metrics. The facts that
1) evaluation of SML under this protocol has outperformed all
previously tested methods [16, 17], and 2) SML is relatively
straightforward to implement, place SML as a natural benchmark against which to compare other algorithms. There are,
nevertheless, a number of questions that remain open with respect to the evaluation of semantic labeling and retrieval systems.
First, because Corel5K is a relatively small database, it is
not clear that the conclusions based on the protocol of [12] will
hold for larger databases, or databases annotated with larger
vocabularies. Second, although relying on a simple image representation, SML requires the specification of a number of parameters (dimension of the feature space, number of mixture
components, number of hierarchical stages for density estimation, etc.). It would be desirable to measure the sensitivity
of retrieval performance to these parameters. Third, a number
of competing semantic retrieval techniques were never considered in [12, 16, 17], including techniques specifically designed
to achieve scalability in database size. Of particular interest is
the method proposed in [14], that relies on a representation
similar to that of SML, but utilizes an alternative image annotation philosophy, which we refer to as supervised categorybased labeling (SCBL). Although this type of labeling is more
error-prone than that adopted by the protocol of [12], it requires significantly less manual effort for the creation of training sets. A quantification of the loss, with respect to retrieval
performance, of SCBL is currently not available.
This work presents the conclusions of a large-scale evaluation effort that aims to answer most of these questions. We
adopt the SML retrieval model and start by testing its scalability, by evaluating it on two databases composed of 30,000 and
60,000 images. We then perform a fairly extensive evaluation
of its robustness with respect to various parameters. This evaluation is followed with a detailed comparison with the model
of [14]. The main conclusions of the study are as follows.
First, the results obtained with the protocol of [12] on Corel5K
seem to hold for larger databases. While precision-recall decreases with vocabulary size (larger vocabularies have more
specialized words, and more words with the same semantic
meaning), annotation performance is qualitatively similar to
that observed on Corel5K. Second, the SML model is quite
robust to the selection of its parameters. We found the performance to be nearly constant over a wide range of config-
urations. Third, the performance loss of SCBL can be quite
significant. Finally, we compare the image representation of
SML with the more sophisticated 2D-MHMM model of [14],
and show that the former has significantly better performance.
This illustrates the important point that more complex models
are not always best, since the added complexity can compromise the generalization ability of retrieval algorithms.
Overall, we see as major contribution of this work, especially when considered as a complement to [16, 12], the unification of results of different annotation systems, previously
only evaluated under different experimental protocols. As far
as we are aware of, this is the first effort of this magnitude:
some of the experiments reported in the following sections required more than 30 hours on a cluster of 2,000 state-of-the-art
Linux processors. We hope that the unified view on the performance of existing annotation algorithms will 1) be of use
for other researchers, and 2) help solidify the current trend,
initiated by [12], towards the adoption of standard databases
and evaluation protocols. The paper is organized as follows.
In Section 2, we start by reviewing, and comparing, the SML
and SCBL models. Next, we discuss image databases, experimental protocol, and details of experiment implementation in
Section 3. Finally, we analyze results in Section 4.
2 Semantic Image Annotation for Large Databases
The goal of semantic image annotation is to, given an image
I, select a set of keywords w from the semantic vocabulary
L = {w1 , . . . , wL } that best describes I. Likewise, the goal
of semantic retrieval is to, given a keyword w, select images
from an image database that contain the associated visual concept. In both cases, learning is based on a training set of
image-caption pairs, D = {(I1 , w1 ), . . . , (ID , wD )}. We
next review two annotations methods, SML and SCBL, that
scale well with database and vocabulary size.
2.1 Supervised multi-class labeling
SML formulates both annotation and retrieval as classification
problems. In particular, SML explicitly makes the elements
of the semantic vocabulary the classes of an M -ary classification problem by introducing: 1) a random vector X of visual features; 2) a random variable W , which takes values
in {1, . . . , L}, so that W = i if and only if x is a sample
from the concept wi ; 3) a set of class-conditional distributions
PX|W (x|i), i ∈ {1, . . . , L} that model the distribution of visual features given the semantic class; and 4) a set of prior
probabilities on the semantic class, PW (i).
Using well known results in statistical decision theory [18],
it is not difficult to show that both labeling and retrieval can
be implemented with minimum probability of error if the posterior probabilities
PW |X (i|x) =
PX|W (x|i)PW (i)
PX (x)
(1)
are available. For annotation, the minimum probability of error rule is to, given a set of query feature vectors x, pick con-
cept
∗
i (x) = argmax PW |X (i|x) = argmax PX|W (x|i)PW (i).
i
i
(2)
For semantic retrieval, given concept wi , the optimal rule is to
select the database image of index
j ∗ (wi ) = argmax PX|W (xj |i)
(3)
j
where xj is the set of feature vectors extracted from the j th
database image Ij . In both cases, the ordering by decreasing
posterior probability is a minimum probability of error ranking
for the remaining keywords or images. Evaluation under the
protocol of [12] has shown that SML outperforms a number of
state-of-the-art unsupervised learning techniques.
2.2 Supervised category-based labeling
SCBL is specifically designed for labeling very large
databases. Because it is time-intensive to hand-label individual images, it proposes the alternative of grouping images into
disjoint categories and labeling all images in each category
with a unique set of ground-truth words, that best describe the
category as a whole. In [14], this idea is implemented in two
steps. First, a classifier is applied to each test image to obtain the top five categories for that image, and the annotations
from those categories are pooled into a list (with frequency
counts for reoccurring annotations). The candidate annotations are then ordered using a statistical test for the hypothesis
that an annotation has entered the list by chance. Specifically,
the probability that the candidate word appears at least j times
in k randomly selected categories is computed
m n−m
k
X
i
k−i
P (j, k) =
I(i ≤ m)
(4)
n
i=j
k
where I(·) is the indicator function, n is the total number of
image categories, and m is the number of image categories
containing the word. A small value of P (j, k) indicates a
low probability that the annotation occurred randomly (i.e. the
word has high significance as an annotation), and hence lower
values of P (j, k) are better.
2.3 Comparison of annotation styles
In comparison with SML, SCBL has two advantages: 1) the
amount of data required from each image category is small
compared to the number of words learned (e.g. in [14], each
image category contains 100 images and is labeled with an average of 4 words); and 2) expanding the database with a new
image category does not require relearning the existing categories, i.e. only the new image category must be learned.
These two properties allow easy addition of new words, or enhancement of previous words, by simply adding new image
categories to the database. In contrast, SML annotation becomes less tractable as the database is expanded in the sense
that: 1) common annotations, such as “landscape”, are associated with many images, and hence learning these class densities becomes time intensive; and 2) when adding new images
to the database, the class mixtures must be relearned for all the
annotations associated with the new image1 . In other words,
not only is the learning problem more difficult because common classes may be associated with a very large number of
images, but expanding an existing database requires relearning all the affected annotation classes.
One disadvantage of SCBL is that a representation of the
actual annotation is never learned. Instead, the system learns
a representation of groups of words, and uses a hypothesis test
to determine the probability that a single word describes the
test image. This is in contrast to SML, which learns a representation of a word directly from all its associated ground truth
images, and is therefore optimal in the minimum probability
of error sense. A second disadvantage of SCBL is that the hypothesis test tends to prefer unique over popular words. As a
result, SCBL may not use generic annotations (e.g. landscape,
water, grass, sky) that may be appropriate for many images.
Also note that P (j, k) tends to take a discrete number of values and that there is no natural ordering for words with the
same value of P (j, k). In contrast, ordering of annotations in
SML is based on posterior probabilities that are very unlikely
to take the same values, and there is almost always a strict
ordering of words.
2.4 Density estimation
The estimation of class conditional densities raises an interesting complexity question: if the database is large, the direct
estimation of PX|W (x|i) from the set of all feature vectors extracted from all images that contain the concept wi is usually
infeasible. One solution is to discard part of the data, but this
is suboptimal in the sense that important training cases may
be lost. An efficient alternative is to adopt the hierarchical
density estimation method of [19, 16, 17], based on a mixture
hierarchy where children densities consist of different combinations of subsets of their parents’ components. With this
method, a class conditional mixture density is learned as follows. First, a mixture density is learned for the feature vectors
extracted from each image, resulting in a mixture density for
each image. Second, a class conditional density is learned by
applying a hierarchical EM algorithm [19] to the densities of
the images.
For classes with a large number of images, multiple hierarchical levels can be used by splitting the set of densities at the
current-level into subsets, and then learning children-densities
from these subsets, and so on. In this way, the class conditional distribution can be learned with logarithmic complexity
on the number of training images. Note that the number of parameters in each image mixture is orders of magnitude smaller
than the number of feature vectors in the image itself. Hence
the complexity of estimating the class mixture parameters is
negligible when compared to that of estimating the individual
mixture parameters for all images in the class. It follows that
the overall training complexity is dominated by the latter.
One final interesting property of the hierarchical method is
that it enforces a data-driven form of regularization which im1 Note, however, that this computation is greatly reduced if the hierarchical
density estimates described later are used, and the image mixtures are saved.
proves generalization [19]. We have observed that, due to this
property, hierarchical class density estimates are much more
reliable than those obtained by direct learning based on standard EM.
recall (NZR), i.e. words with |wc | > 0. This can be seen as
the number of words the system has effectively learned.
The performance of semantic retrieval is also evaluated by
measuring precision and recall. Given a query term and the
top n image matches retrieved from the database, recall is the
percentage of all relevant images contained in the retrieved
3 Experimental Evaluation
set, and precision the percentage of the n which are relevant
Recent works [11, 12, 15] have adopted a “de-facto” experi- (where relevant means that the ground-truth annotation of the
mental protocol, which we refer to as Corel5K, for evaluating image contains the query term). Under the protocol of [12], resemantic annotation systems. This protocol has two signif- trieval performance is evaluated by the mean average precision
icant limitations. First, because the database on which it is (MAP). The retrieved images are ordered according to their
based is relatively small, it relies on a very small number of scores, and the average precision is computed by averaging
examples for many of the semantic labels. This makes it dif- over only the precision values where the recall changes (i.e.
ficult to guarantee that the resulting annotation systems have where relevant items occurred), while increasing the number
good generalization. Second, because the concept vocabulary of retrieved images. The MAP is the mean of the average preis also relatively small, it does not necessarily test the scala- cision values for all the query words.
bility of annotation/retrieval algorithms. Some of these limiThe Corel30K protocol is similar to Corel5K but substantations are corrected by the Corel30K protocol, an extension
tially larger, containing 31,695 images and 5,587 words. Of
of Corel5K based on a substantially larger database. None of
the 31,695 images, 90% were used for training (28,525 imthese protocols has, however, been used to evaluate some of
ages) and 10% for testing (3,170). Only the words (950 in tothe most scalable techniques available in the literature. One
tal) that were used as annotations for at least 10 images were
alternative protocol, originally proposed by [14], is more suitlearned. Corel30K is much richer than Corel5K, in terms of
able to test such systems. We refer to it as PSU, and review
number of examples per label and database size, therefore posthe three protocols in this section.
ing a much stronger scalability challenge.
3.1 Corel5K and Corel30K
The evaluation of a semantic annotation/labeling and retrieval
system requires three components: an image database with
manually produced annotations, a strategy to train and test the
system, and a set of measures of retrieval and annotation performance. The Corel5K benchmark is based on the Corel image database [11, 12, 15]: 5,000 images from 50 Corel Stock
Photo CDs, were divided into a training set of 4000 images,
a validation set of 500 images, and a test set of 500 images.
An initial set of model parameters is learned on the training
set. Parameters that require cross-validation are then optimized on the validation set, after which this set is merged with
the training set to build a new training set of images. Noncross-validated parameters are then tuned with this training
set. Each image has a caption of 1-5 keywords, and there are
371 keywords in the entire data set (the test set only contains
280 of these).
Image annotation performance is evaluated by comparing
the captions automatically generated for the test set, with
human-produced ground-truth. Similarly to [12, 15] we define
the automatic annotation as the five semantic classes of largest
posterior probability, and compute the recall and precision of
every word in the test set. For a given semantic descriptor, assuming that there are |wH | human annotated images in the test
set, and the system annotates |wa |, of which |wc | are correct,
recall and precision are given by:
3.2 PSU protocol
PSU is also based on the Corel set, and contains 60,000 images
[14] with 442 annotations. Unlike Corel5K and Corel30K, the
PSU ground truth was not obtained by annotating each image
separately. Instead, images were associated with an image category (the image set was split into 600 image categories with
100 images each), which were then annotated with a general
description that reflects the image category as a whole, but
may not accurately characterize each individual image. This
strategy saved human annotation time but produced somewhat
noisy annotations. For example, some images from the “tiger”
category, which is labeled with “tiger”, “grass”, and “trees”,
do not contain “trees”. For performance evaluation, 40% of
the PSU images were reserved for training (23,878 images),
and the remainder (35,817 images) used for testing. Note
that [14] only used 4,630 of the 35,817 possible test images,
whereas the experiments reported here are based on the entire
test set.
Following [14], the performance of the image category classifier is evaluated by its accuracy, where a test image is considered to be categorized correctly if any of the top r categories is the true category. Also following [14], the performance of SCBL annotation is measured with the “mean cover|wc |
|wc |
age”, which is the percentage of ground-truth annotations that
,
precision
=
(5)
recall = |w
|wa |
H|
match the computer annotations. In addition, we also evaluAs suggested by [12, 15], the values of recall and precision ate the annotation and retrieval performance with the standard
are averaged over the set of words that appear in the test set. measures used in Corel5K and Corel30K. A summary of the
Finally, we also consider the number of words with non-zero experimental protocols is given in Table 1.
images
5,000
31,695
59,695
labels
260
950
442
training set
4,500 (90%)
28,525 (90%)
23,878 (40%)
testing set
500 (10%)
3,170 (10%)
35,817 (60%)
Table 1: Image databases and experimental setup
Class
Representation
GMM-DCT
2D-MHMM [14]
r=1
r=2
r=3
r=4
r=5
0.209
0.119
0.270
0.171
0.309
0.208
0.338
0.232
0.362
0.261
0.4
0.35
Categorization Accuracy
DB name
Corel5K
Corel30K
PSU
0.3
0.25
0.2
0.15
0.05
Table 2: Accuracy of image categorization on PSU database.
A image is correctly categorized if any of the top r categories
is the true category.
3.3 Image representation
All experiments were based on the image representation of
[16, 17]. In particular, we use the YBR color space, and each
image was decomposed into a set of overlapping 8 x 8 windows, which were extracted with a sliding window that moved
by one pixel between consecutive samples. A feature vector
was obtained, from each location of the three color channels,
by the application of the discrete cosine transform (DCT). The
image was then represented as a bag X = {x1 , · · · , xN } of independently sampled feature vectors.
3.4 Implementation details
For the implementation of SML, a Gaussian mixture was fit
to the set of images associated with a given annotation using
the hierarchical method discussed in Section 2.4. A 2-level
hierarchy was constructed by first learning mixtures for the
feature vectors of the images, and then learning the class conditional from the image mixtures. For the 3-level hierarchy, an
additional level split the image mixtures into groups of 250,
and mixtures were learned for those groups. The class conditional density was finally learned from the group mixtures.
We refer to this representation as GMM-DCT. Unless otherwise noted, the image densities used 8 mixture components
and the class conditionals used 64 components. The Gaussian components had diagonal covariance matrices and used
the full 192-dimensional space.
Using SML, test images were annotated with the five labels of largest posterior probability. For single query image
retrieval (i.e. ranked retrieval), each image was first annotated
with five labels. When retrieving images for a query word, all
images that received the word as an annotation were selected,
and ordered based on the posterior probability of word given
image.
The implementation of SCBL for the PSU database used
the GMM-DCT representation for each image category. Class
conditional distributions also used 64 mixture components and
the full feature space. Experiments were conducted on a cluster of 2,000 state-of-the-art Linux machines, and ran between
1 hour for Corel5K and 34 hours for PSU. These times are
overestimates, since the experiment was sometimes preempted
r=1
r=2
r=3
r=4
r=5
0.1
0
20
40
60
80
100
120
140
Dimension of Feature Space
160
180
200
Figure 1: Accuracy of image categorization on PSU using
GMM-DCT versus the dimension of the DCT feature space.
by other jobs on the cluster.
4 Experimental Results
In this section we present the results of annotation experiments. We first compare the performance of SML and SCBL,
and then present robustness results of SML on the three
databases.
4.1 Comparison of SML and SCBL
We started by comparing the image categorization performance between the GMM-DCT class representation and the
representation of [14]. In [14], an image category is represented by a two-dimensional multi-resolution hidden Markov
model (2D-MHMM) defined on a feature space of localized
color and wavelet texture features at multiple scales. Table 2
shows the accuracy of image categorization using the two class
representations. GMM-DCT outperformed the 2D-MHMM in
all cases, with an improvement of about 0.10 (from 0.26 to
0.36). Figure 1 shows the categorization accuracy of GMMDCT versus the dimension of the DCT feature space. It can
be seen that the categorization accuracy increases with the dimension of the feature space, but remains fairly stable over a
significant range of dimensions.
We next compared the annotation performance of SCBL,
using GMM-DCT and the 2D-MHMM (we denote the combinations by SCBL-GMM-DCT and SCBL-2D-MHMM). The
images were annotated by applying a threshold of 0.0649 to
P (j, k), as described in [14]. Table 3 shows the mean coverage of SCBL-GMM-DCT and SCBL-2D-MHMM with and
without thresholding. GMM-DCT annotations outperformed
those of 2D-MHMM by about 0.12 (from 0.22 to 0.34 with
threshold, and 0.47 to 0.61 without). Figure 2 shows the
mean coverage versus the dimension of the DCT feature space.
Again, performance increases with feature space dimension,
but remains fairly stable over a large range of dimensions.
Finally, we compared SCBL and SML when both methods
used the GMM-DCT representation. SCBL annotation was
performed by thresholding the hypothesis test (SCBL-GMMDCT threshold), or by selecting a fixed number annotations
Method
SCBL-GMM-DCT
SCBL-2D-MHMM [14]
threshold=0.0649
0.3420
0.2163
no threshold
0.6124
0.4748
0.24
SML−GMM−DCT, 2−level
SML−GMM−DCT, 3−level
SCBL−GMM−DCT, threshold
SCBL−GMM−DCT, fixed
0.22
0.2
Table 3: Mean coverage of annotation on PSU database using
SCBL.
Precision
0.18
0.16
0.14
0.12
GMM−DCT no threshold
GMM−DCT th=0.0649
2D−MHMM no threshold
2D−MHMM th=0.0649
0.7
0.1
0.08
Mean Coverage
0.6
0.06
0.04
0.5
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Recall
0.4
Figure 3: Precision-Recall for SCBL and SML using GMMDCT on the PSU database.
0.3
0.2
annotation ground truth.
0
50
100
150
Dimension of Feature Space
200
Figure 2: Mean coverage of annotation of PSU using SCBLGMM-DCT versus the dimension of the DCT feature space.
(SCBL-GMM-DCT fixed). SML classifiers were learned using both 2-level and 3-level hierarchies. Figure 3 presents
the precision-recall (PR) curves produced by the two methods.
Note that SML trained with the 3-level hierarchy outperforms
the 2-level hierarchy. This is evidence that the hierarchical
EM algorithm provides some regularization of the density estimates, which improves the performance.
The SML curve has the best overall precision at 0.236, and
its precision is clearly superior to that of SCBL at most levels of recall. There are, however, some levels where SCBLGMM-DCT leads to a better precision. This is due to the
coupling of words within the same image category, and to
the noise in the ground truth annotations of PSU. Note that
if the correct category is in the top 5 classified categories, then
the list of candidate words will contain all of the ground truth
words for that image. Eventually, as the image is annotated
with more words from the candidate list, these ground truth
words will be included, regardless of whether the ground truth
actually applies to the image (i.e. when the ground truth is
noisy). In result, recall and precision are artificially inflated
as the number of annotations increases. On the other hand,
for SML, each word class is learned separately from the other
words. Hence, images will not be annotated with the noisy
word if the concept is not present, and the precision and recall
can suffer. Finally, for SCBL-threshold, the PR curve has an
unusual shape. This is an artifact that arises from thresholding
a hypothesis test that has discrete levels.
In summary, the experimental results show that the
GMM-DCT representation substantially outperforms the 2DMHMM of [14] in both image categorization and annotation using SCBL. When comparing SML and SCBL based on
the GMM-DCT representation, SML achieves the best overall
precision, but for some recall levels SCBL can achieve a better
precision due to coupling of annotation words and noise in the
4.2 Robustness and scalability of SML
We have already seen that, under the SCBL model, both the
categorization and annotation performance of the GMM-DCT
representation are fairly stable with respect to the number of
mixture components and feature space dimension. We now
report on experiments performed to evaluate the robustness of
SML-GMM-DCT to these parameters. Figure 4 (left) shows
the PR curves obtained for annotation on Corel5K, as a function of the number of mixture components used to model class
conditional densities. Note that the PR curve remains fairly
stable above 64 components. Figure 4 (right) shows the PR
curve for annotation with 64 components while varying the
feature space dimension. In this case, stability is achieved
above 63 dimensions.
To test scalability, the SML annotation experiment was repeated on the larger Corel30K. Figure 5 shows the performance obtained with 64 and 128 mixture components, learned
with either the 2-level or 3-level hierarchy. Annotation performance on the larger database is qualitatively similar to that
obtained on the smaller Corel5K database (e.g. compare the
shape of the PR curves with those of Figure 4 (left)), albeit
with overall lower precision and recall levels. This is due
to the difficulty of learning specialized annotations, and to
the presence of different annotations with the same semantic meaning, which are both more frequent on Corel30K. In
addition, a 3-level hierarchy outperforms the standard 2-level
hierarchy for both 64 and 128 components. This indicates that
the regularization of the 3-level structure is superior to that
of the standard hierarchical organization. The differences are
nevertheless not staggering, suggesting some robustness with
respect to this parameter.
Overall, these experiments indicate that 1) SML is fairly
stable with respect to its parameter settings, and 2) results on
Corel5K are a good indication of the relative performance of
different techniques (albeit the absolute PR values are likely to
be overestimated). Finally, Table 4 shows the average precision and recall for semantic annotation with five labels using
SML with 64 mixture components and the full DCT space.
0.25
0.24
C=8
C = 16
C = 32
C = 64
C = 128
C = 256
0.22
0.2
0.2
0.16
Precision
Precision
0.18
d=3
d=9
d = 18
d = 63
d = 147
d = 192
0.14
0.12
0.15
0.1
0.1
0.05
0.08
0.06
0
0.04
0.1
0.2
0.3
0.4
Recall
0.5
0.6
0.7
0
0.1
0.2
0.3
0.4
Recall
0.5
0.6
0.7
Figure 4: Precision-recall curves for annotation on Corel5K using SML while varying: (left) the number of mixture components
(C); (right) the dimension of the DCT feature space (d) for 64 mixture components.
0.16
64 components, 2−level
64 components, 3−level
128 components, 2−level
128 components, 3−level
0.14
Precision
0.12
0.1
0.08
0.06
0.04
0.05
0.1
0.15
0.2
0.25
0.3
Recall
0.35
0.4
0.45
0.5
Figure 5: Precision-recall curves for annotation on Corel30K
using SML-GMM-DCT.
Figure 6 shows several examples of SML on Corel30K. The
automatic annotations are plausible, even though not perfectly
matching the human ground-truth.
4.3 Ranked retrieval results
Figure 7 presents results of ranked retrieval on Corel5K for
different numbers of mixture components and DCT dimensions. The left plot depicts the MAP for all 260 words, while
the one in the center shows the same curves for words with
non-zero recall. In both cases, the MAP increases with the
number of mixture components, stabilizing above 128 components. The plot on the right shows the number of words with
non-zero recall, which decreases with the number of mixture
components, once again stabilizing above 128 components.
Table 4 presents retrieval results obtained with SML on the
the three databases.
5 Conclusions
In this work we have presented a comparison of the SML and
SCBL models for semantic annotation and retrieval on very
large databases. Various conclusions can be taken from our
experiments. First, the GMM-DCT classifier outperforms the
2D-MHMM classifier of [14] in both image categorization and
SCBL image annotation. When using GMM-DCT, SML has
better overall precision than SCBL, although the latter may
have some values of precision and recall that are better when
the test image is annotated with a substantial number of words.
This is due to the noisy ground-truth labels, and label coupling
within image categories. Second, the performance of SML is
robust to changes in the dimension of the feature space, and
number of mixture components. Finally, SML seems to be
scalable to large databases, where annotation performance displays similar characteristics to that of smaller databases. For
large databases, using a 3-level hierarchical method for learning the class mixtures can improve the generalization of the
system, as compared to the standard 2-level hierarchy. The
differences are however not staggering, confirming the robustness of SML to its parameter settings.
Acknowledgments
The authors would like to thank Kobus Barnard for providing
the Corel dataset used in [11], David Forsyth for providing the
Corel30K dataset, James Wang for the PSU dataset used in
[14], Google Inc. for providing the computer resources for the
experiments, and Gustavo Carneiro for helpful conversations.
This research was partially supported by NSF CAREER award
IIS-0448609 and a grant from Google Inc.
References
[1] A. W. M. Smeulders, M. Worring, S. Santini, A. Gupta,
R. Jain. Content-based image retrieval: the end of the
early years. PAMI, 22(12):1349-80, 2000.
[2] R. Picard. Digital libraries: meeting place for high-level
and low-level vision. In Proc. Asian Conf. on Computer
Vision, Dec. 1995.
[3] M. Szummer and R. Picard. Indoor-outdoor image classification. In Workshop in Content-based access to image
and video databases, 1998.
Database
Corel5K
Corel30K
PSU
Precision
0.2163
0.1261
0.1519
Recall
0.2871
0.2122
0.3158
Semantic Annotation
Precision NZR Recall NZR
0.4499
0.5972
0.2826
0.4756
0.1626
0.3379
# of NZR words
125
424
413
Ranked Retrieval
MAP MAP NZR
0.3023
0.6288
0.2113
0.4734
0.2552
0.2732
Table 4: Results for semantic annotation using SML-GMM-DCT: average precision and recall for all words, and words with
non-zero recall (NZR). Results for ranked retrieval: mean average precision (MAP) for all words, and for words with non-zero
recall.
Human
SML
Human
SML
Human
SML
sky, track, rail
rail, train, tracks,
harbor, boats
cowboys, horses,
water, people
cowboys, horses, fields,
landscape, fence
boats, water,
people, carnival
coast, waves,
harbor, shore, glacier
castle, building,
water, reflection
palace, mosque,
sculpture, castle, arch
musician, people
fireworks, night,
sky, burst
burst, fireworks,
quartz, gem, lights
butterfly, side, leaves
butterfly, wings, fungus,
mushroom, branch
sailboards, people,
water, waves
sailboard, sea,
waves, shore, canoe
plants, market, model,
costumes, costume
man, donkey,
people, building
castle, tower, sculpture,
landmark, statue
valley, grass, sky
desert, canyon, pyramid,
valley, formation
monkey, leaf, zoo
model, zoo, branch
monkey, rainbow
food, meal, cuisine
cuisine, dessert,
meal, snack, sunset
Figure 6: Examples of semantic annotation using SML-GMM-DCT on the Corel30K database.
0.3
0.25
0.2
0.15
0.1
0
50
100
150
Dimension of Feature Space
200
0.7
150
0.65
140
Number of NZR words
Mean Average Precision (NZR words)
Mean Average Precision (all 260 words)
0.35
0.6
0.55
0.5
0.45
0.4
0.35
130
120
110
100
C=8
C = 16
C = 32
C = 64
C = 128
C = 256
90
80
0
50
100
150
Dimension of Feature Space
200
70
0
50
100
150
Dimension of Feature Space
200
Figure 7: Ranked retrieval on Corel 5k using SML with different mixture components (C): (left) MAP for all the words;
(middle) MAP for words with non-zero recall; (right) number of words with non-zero recall.
[4] A. Vailaya, A. Jain, and H. J. Zhang. On image classification: city vs. landscape. Pattern Recognition, vol. 31,
pp. 1921-36, 1998.
[5] N. Haering, Z. Myles, and N. Lobo. Locating deciduous
trees. In Workshop in Content-based Access to Image
and Video Libraries, pp. 18-25, 1997.
[6] D. Forsyth and M. Fleck. Body plans. In Proc. CVPR,
1997.
[7] Y. Li and L. Shaprio. Consistent line clusters for building
recognition in CBIR. In Proc. ICPR, vol. 3, pp. 952-956,
2002.
[8] K. Barnard, P. Duygulu, N. de Freitas, D. Forsyth, D.
Blei, and M. I. Jordan. Matching words and pictures.
JMLR, 3:1107-1135, 2003.
[9] D. Blei and M. Jordan. Modeling annotated data. In
Proc. ACM SIGIR, 2003.
[10] P. Carbonetto, N. de Freitas, and K. Barnard. A statistical
model for general contextual object recognition. In Proc
ECCV, 2004.
[11] P. Duygulu, K. Barnard, N. Freitas, and D. Forsyth. Object recognition as machine translation: learning a lexicon for a fixed image vocabulary. In Proc. ECCV, 2002.
[12] S. Feng, R. Manmatha, and V. Lavrenko. Multiple
Bernoulli relevance models for image and video annotation. In Proc. CVPR, 2004.
[13] P. Carbonetto, H. Kueck, and N. Freitas. A constrained
semi-supervised learning approach to data association.
In Proc. ECCV, 2004.
[14] J. Li and J. Z. Wang. Automatic linguistic indexing of
pictures by a statistical approach. PAMI, 2003.
[15] V. Lavrenko, R. Manmatha, and J. Jeon. A model for
learning the semantics of pictures. In NIPS, 2003.
[16] G. Carneiro and N. Vasconcelos. A database centric view
of semantic image annotation and retrieval. In Proc.
ACM SIGIR, 2005.
[17] G. Carneiro and N. Vasconcelos. Formulating semantic
image annotation as a supervised learning problem. In
Proc. CVPR, 2005.
[18] R. Duda, P. Hart, and D. Stork. Pattern Classification.
John Wiley and Sons, 2001.
[19] N. Vasconcelos. Image indexing with mixture hierarchies. In Proc. CVPR, 2001.